基于BP神经网络自学习的双目视觉系统的研发

在摄像机标定的过程中，深度信息的丢失，摄像机镜头的畸变以及图像处理时误差等因素都影响标定的精度．本论文采用BP神经网络的自学习的性能，开发出一套双目视觉系统．以匹配点在左、右图像的坐标为网络的四路输入，通过网络得到三路输出，性能指标为该对应点在世界坐标系的坐标和网络输出的差值的平方和，根据梯度下降法来调整各神经元之间的连接权值，求得网络达到给定的误差时的各节点问权值．这样，双目视觉系统两个摄像机的投影矩阵可以用神经网络的权值与激发函数来代替，完成系统的标定．最后对系统进行精度分析．

Development Of Binocular Vision System Based On BP Neural Network Self-learning

Those factors such as the loss of depth information,distortion of camera lens,and errors caused by image processing,influence precision of camera calibration.In this paper binocular vision system is developed by means of back propagation neural network with self-learning performance.There are 4 inputs,i.e.,the coordinates of image of a match point in left and right camera,3 outputs in the network,and the sum square of errors between the outputs of the network and actual coordinates measured in world system is taken as performance index.The network weights are tuned in the light of gradient descend method and can be achieved until the given sum square of errors is least.Thus each projection matrix of two cameras of binocular vision system can be replaced by the weights and the activation function of the neural network,and calibration of system is finished.Finally,the precision analysis is carried out for the system.